Did the rediscovery discount itself cause the collapse in finds?

EXPERIMENTS.md · lines 8786–8819

E144 — The novelty pressure deleted the findings instead of redirecting the search

E140 ran the fly on the whole brain with novelty in the reward. Against E136 (mushroom body only, no novelty):

E136   fly, 5311 neurons, no novelty    AUC_Q 1.00   distinct 5.00   best -74
E140   fly, whole brain, novelty        AUC_Q 0.43   distinct 1.00   best -27

Prediction 1 is falsified and then some — I predicted distinct finds would rise well above 5.0 and they fell to 1.0. Prediction 2 held (quality fell, as the trade intended). Prediction 3 is untestable on one find.

The cause is a bug I introduced an hour earlier, not a property of the fly or the brain. trial.run uses a single p for two different questions: what the generator is paid, and whether a composition counts as a find. The novelty factor multiplied that one number, so:

drive -80 meV/atom  ->  p 0.961  ->  x0.25 for being near a published alloy  ->  0.240
qualifying bar                                                                  0.832

An excellent composition next to a known alloy was paid 0.24 and therefore could never qualify. Almost everything good is near a known alloy, so almost nothing qualified. The novelty pressure did not push the search outward; it erased the results.

Fixed: verify() now returns p (the ladder's verdict, unpenalised) and p_reward (what the generator is paid, penalised). trial.run teaches on p_reward and counts qualifiers on p.

And E140 was a confounded experiment regardless. It changed two things at once — the substrate went from 5,311 neurons to 164,506 and novelty entered the reward — so even without the bug it could not have attributed the difference. The comparison has to be run one change at a time, and that is now three runs: whole brain without novelty, mushroom body with novelty, and the corrected both.

No conclusion about the whole brain can be drawn from E140. Its result is a measurement of my own defect.